Financial Security Institute Establishes Security Evaluation Standards for AI Agents in the Financial Sector

The Financial Security Institute announced today (October 7) that it has established the "Financial Sector AI Agent Security Evaluation Standards" to inspect new security risks arising from the expansion of artificial intelligence (AI) agents in the financial sector.

An AI agent is an artificial intelligence system that independently formulates plans to achieve given goals and utilizes tools and external systems to perform actual tasks.

Because they handle everything from problem definition to execution autonomously without human intervention, security threats are more likely to have a direct impact on actual operations compared to traditional AI.

The Financial Security Institute analyzed major domestic and international AI agent security threats and countermeasures to derive inspection items, deciding to reflect them in the "AI Red Teaming" standards.

AI Red Teaming is a mock hacking technique that proactively attacks from a hacker's perspective to discover vulnerabilities and check the adequacy of responses.

The evaluation criteria are divided into 17 items across 5 categories: ▲ response to data poisoning attacks, ▲ response to model poisoning attacks, ▲ response to data and model extraction attacks, ▲ response to evasion attacks, and ▲ AI agent security.

Among these, 10 items prepare for threats newly generated by the adoption of AI agents.

The identified AI agent security threats, actual attack techniques, and final evaluation criteria are scheduled to be released in this year's "AI Red Team Report."

Park Sang-won, President of the institute, stated, "With the establishment of these evaluation standards, we will practically verify the risks of AI agents in the financial sector and actively support financial institutions in utilizing AI technology safely and reliably."

(Photo provided by Financial Security Institute, Yonhap News)
※ Please note: This article was translated by AI and may contain errors.